---
title: "Agogic: Performance-Timed Music Tokens for LLM-Native Text-to-Symbolic-Music Generation"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.03999v1"
permalink: "https://feed.astralor.com/items/itm_256fee446ba5ddb924ccd4b6"
published: "2026-08-04T17:56:49.000Z"
score: 77
channel: "ai"
tags: ["PMT","Qwen3.5","Text-to-Music Generation","Symbolic music","Music Tokenization"]
---

# Agogic: Performance-Timed Music Tokens for LLM-Native Text-to-Symbolic-Music Generation

## AI commentary

> The real contribution is not another SOTA score, but turning 'does tokenization matter?' into a measurable fact. Teams working on music AIGC should try a representation change before scaling up further, and evaluation-focused readers will find the released corpora and diagnostic immediately useful.

## At a glance

This paper isolates tokenization from all other variables by fixing pretrained Qwen3.5 backbones (0.8B–27B), data, budget, and decoding, then swapping only the music representation across seven tokenizations. The results are striking: representation, not model size, is the binding variable for distributional fidelity. Scaling the backbone 34x barely moves Fréchet Music Distance (FMD), while changing representation halves it. PMT, a performance-resolution stream with 10 ms timing, per-note velocity, multi-track texture, and 609 symbols, reaches FMD 159 at 0.8B, beating beat-grid tokenizations at 27B. The effect persists on a from-scratch 26M model and a second tokenizer, indicating it is a property of the class. Caption adherence is weak but separable; a lightweight decode-time constraint sharply improves instrument and key accuracy. The authors release a harness, 25+ checkpoints, two corpora (86.6k aligned; 6.25M captioned), and a diagnostic showing published systems largely ignore captions.

## Key points

- With all else fixed, swapping only music representation matters more than model size: 34x scaling barely changes FMD, while representation change halves it.
- PMT (10ms timing, per-note velocity, multi-track, 609 symbols) achieves FMD 159 at 0.8B, beating a 27B beat-grid model and replicating on a 26M from-scratch backbone.
- Native caption adherence is weak but separable: a light decode-time constraint lifts instrument F1 from .28 to .60 and Correct-Key from .16 to .35 with no distributional cost.
- They release the eval harness, 25+ checkpoints, an 86.6k aligned corpus, and a 6.25M captioned corpus, plus an imprinting diagnostic for future representation claims.

## Full text

Text-to-music language models begin with a choice usually made by default: how to tokenize music. Normally entangled with backbone, data, and recipe, its effect has never been measured in isolation. We fix pretrained Qwen3.5 (0.8B-27B), data, budget, and decoding, and swap only the representation across seven tokenizations, anchoring texture metrics to each representation's model-free ceiling. The ordering is clean and surprising: representation, not model size, is the binding variable for distributional fidelity. Scaling the backbone 34x barely moves Frechet Music Distance (FMD), whereas switching representation halves it. PMT, a performance-resolution stream we release (10 ms timing, per-note velocity, multi-track texture; 609 symbols), reaches FMD 159 at 0.8B against 272-286 for beat grids (1.7-1.8x lower, up to 2.8x elsewhere; non-overlapping bootstrap CIs), so a 0.8B performance-resolution model beats a 27B beat grid. It reappears on a 26M from-scratch backbone and a second performance-resolution tokenizer: a property of the class, not one lucky vocabulary. Nor is it a finer-lattice artifact: snapping PMT's onsets to the beat grids' resolution still leaves it 67-129 FMD ahead of both (n=500). The effect is distributional; whether it is audible is a separate question, left open by our probe, with a human study pre-registered. Native caption adherence is weak but separable: a lightweight decode-time constraint doubles instrument-F1 (.28 to .60) and Correct-Key (.16 to .35) at no distributional cost. We release the harness, 25+ checkpoints, two corpora (86.6k aligned across caption/MIDI/ABC/audio; 6.25M captioned, the largest for music), and an imprinting diagnostic: published text-to-MIDI systems reproduce their training distribution near-invariant to the caption (72% vs. 71% chord-time on disjoint domains). The field's next representation claim can now be measured, not asserted.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_256fee446ba5ddb924ccd4b6